Research Before Outreach

What to research, and when to stop

Research has one job: to find a reason this account and this person might care now. Three layers are worth checking, in order.

The account. Does it fit your ideal customer profile (Lesson 1.3)? Is there a trigger: a funding round, an acquisition, a new executive, an expansion, a regulatory change, a public statement of priorities? A trigger is the strongest reason to write, because it ties your message to something the buyer is already thinking about.

The person. What is their role in a purchase like yours? What have they said publicly about the problem you solve? Do you share a connection, a former employer, a customer?

Your history. Have you spoken to this account before? Is there an open or lost deal? Has anyone on your team already written to this person? Sending a cold opener to someone a colleague is mid-conversation with is the most avoidable embarrassment in sales.

Stop when you have one good reason. Salesforce's 2026 State of Sales report (vendor survey) found that sellers expect AI agents to cut prospect research time by 34%; the risk is spending the saved time researching more accounts less carefully. One verified, relevant fact beats five plausible ones.

A verification standard

A fact you put in front of a buyer should pass three checks.

It is quoted from a source you can point to. Not a summary of a summary, not an AI paraphrase: the actual sentence on the actual page, which you or anyone else can go and read.

It is about the right person or company. Two people with the same name at the same large firm are common. Name plus company alone is weak; add a second corroborator such as title, location or the company's own domain.

It is dated, and recent enough to matter. A 'recent post' from two years ago is not recent. A funding round from last year may be the wrong round now.

If a fact fails any check, drop it rather than softening it into a vaguer claim. 'I saw you're expanding' built from a fact you could not verify is still a guess, just a less falsifiable one.

For regulated sellers the standard is not optional. FINRA Rule 2210, which governs broker-dealer communications with the public, requires communications to be fair and balanced and prohibits any false, exaggerated, unwarranted, promissory or misleading statement or claim. A personalized opener that gets a prospect's deal history wrong is a misstatement by the firm, not a typo.

Where the data may come from

How you gather data matters as much as what you gather. LinkedIn's User Agreement (in its version effective November 3, 2025) prohibits developing, supporting or using software, scripts, robots or other means to scrape or copy the service, including profiles. Many prospecting tools do exactly that; using their output can put your own account and your vendor's at risk.

Data protection law also applies to business contacts. In the UK, the Information Commissioner's Office states that UK GDPR applies when you process personal data for direct marketing even in a business context, and that business contacts have an absolute right to object to direct marketing (Lesson 3.4 returns to consent). The safe pattern is to research from public, first-party sources such as the company's own website, press releases, filings and the person's own published writing, to keep a record of where each fact came from, and to stop using a fact when its source no longer supports it.

Prospect research in Kavanah

Kavanah's prospect research is built around the standard above, and it is opt-in. In Settings, under the AI Agent tab's oversight options, 'Research contacts before personalized cadence steps' is off by default; it also requires cadence personalization to be on, because research without a personalized step would buy nothing. When enabled, it researches only contacts whose next cadence step is personalized and due within 48 hours, at most once per contact per 30 days and within a daily cap per workspace, and it spends AI credits (roughly 27 per contact in a typical case). You can also research one contact on demand by asking the agent, which uses the research_contact tool, and read what is stored with get_contact_research.

What it stores is facts, never prose. A candidate fact is kept only if its quote appears verbatim on the page that was fetched, the identity is corroborated by full name, firm and at least one further independent signal, the fact is dated and within an age limit for its type (90 days for a recent post, 180 for news, three years for a deal), and a second, independent check, a fresh fetch plus a separate model judgment that the quote fully supports the statement, agrees. Anything that fails is discarded with a counted reason, not rewritten into something weaker. LinkedIn, Facebook and Instagram pages are never scraped.

When a personalized opener uses research, it may only cite stored facts by their identifier, and the identifiers it used are recorded against that send, which gives you a trail from a sentence in a sent email back to its source and quote. An admin can retract a fact, and a retracted fact stays retracted even if a later run finds it again.